Prometheus:Basics

Logging in

Access to the Prometheus cluster is facilitated by the access host at:

prometheus.cyfronet.pl (alias for the following hosts):

login01.prometheus.cyfronet.pl
login02.prometheus.cyfronet.pl

eg.

ssh <plgtestowy>@prometheus.cyfronet.pl

You can also use the short alias pro.cyfronet.pl

Login node	Description
prometheus.cyfronet.pl pro.cyfronet.pl	Default login node for PLGrid users
prace.pro.cyfronet.pl	Dedicated login node for PRACE users
prace-int.pro.cyfronet.pl	Alias of PRACE login node within PRACE internal network

An SSH client is required for login.

SSH fingerprints:

 (ECDSA)   SHA256:tKk4RpOzWSTf+H8oN7COX1o5xMGSX7vMk956oHcdSB8
 (ED25519) SHA256:roerSF6HBUEqJpoPk36BXwz7NsPRIrTguoApYoYK+ro
 (RSA)     SHA256:v/iUvEh8faHN7IoZ5Vpm0kU6OYKXkGyWp9O9ZFkbeDw

 (ECDSA)   MD5:16:e5:7f:69:45:d0:0f:6c:49:8d:c0:99:0b:e1:e9:dd
 (ED25519) MD5:33:4e:15:cf:71:07:b3:6d:62:0e:b3:7a:9e:7c:7e:13
 (RSA)     MD5:f0:ea:1e:85:79:e0:29:f7:de:64:7a:df:ed:c4:f4:fa

Please ensure you accept a valid key.

Available software

A list of available software can be found here.

Access to software is provided through the modules tool.

Partitions

SLURM replaces PBS queues with the concept of partitions. The following paritions are currently available for Prometheus users:

Parition name	Max task duration	Description
plgrid	72:00:00	Default partition
plgrid-testing	1:00:00	For testing and preparing computations (e.g. building applications, copying files etc.)
plgrid-now	12:00:00	Recommended for interactive jobs, for testing and preparing computations (e.g. building applications, copying files etc.), with maximum limits: 1 node, 24 cores, without GPGPU nodes
plgrid-short	1:00:00	For short jobs (walltime shorter than 1h); provides an increased likelihood that your job will be quickly scheduled for execution
plgrid-long	168:00:00	For long jobs (as specified in the relevant grant)
plgrid-large	24:00:00	For jobs requiring at least 432 cores (18 full nodes); contains additional resources not available through the plgrid partition
plgrid-gpu	72:00:00	For jobs requiring GPGPU computations (note that access to GPGPU must be allowed under the relevant grant)
plgrid-gpu-v100	12:00:00	For jobs requiring GPGPU V100 computations (note that access to GPGPU must be allowed under the relevant grant)
prace	72:00:00	Partition for PRACE projects

Further information regarding any specific partition may be obtained by running the following command:

scontrol show partition [<partition_name>]

Notes:

• Selecting the right partition and specifying the expected execution time greatly increase the chances that your job will be quickly scheduled for execution. We recommend that jobs be preferentially submitted to partitions with short maximum execution times (e.g. plgrid-short where feasible).
• Failing to specify a partition will cause your job to be submitted to the default partition, i.e. plgrid.
• Jobs submitted to plgrid (or submitted without specifying a partition) may be automatically transferred to plgrid-large or plgrid-short by the queue manager, if this will cause them to be executed faster.
• Currently, all users may access plgrid, plgrid-testing, plgrid-short and plgrid-large. Access to other partitions is regulated by grants and provided on a case-by-case basis in the course of evaluating proposed grants.
• If you wish to access plgrid-long, this must be clearly expressed in your grant application.
• Access to plgrid-gpu requires applying for a separate grant which is wholly dedicated to GPGPU computations.
• plgrid-testing must not be used for large-scale computations. This partition is specifically meant for tests and other maintenance tasks (e.g. building applications, interactively copying data etc.)

Disk storage

Location	Volume quota	Files quota	Purpose	Access	Throughput	Available space	Backups	Support for intensive writes and/or reads	Availability	Remarks
$HOME	40 GB	1 000 000	user home directory	NFS	1 GB/s	∞	yes	no	global
$SCRATCH	100 TB	1 000 000	storage for temporary job data	Lustre	120 GB/s	5 PB	no	yes	global	data older than 30 days will be automatically deleted
$PLG_GROUPS_STORAGE/<team_name>	grant total		Storage for PL-Grid team datasets	Lustre	30 GB/s	2.5 PB	no	no	global
$SCRATCH_LOCAL	768 GB	∞	local scratch storage	Lustre	120 GB/s	-	no	yes	limited to single node	unavailable following completion of job; requires -C localfs

Environmental variables

Personal:

• $HOME, $PLG_USER_STORAGE, eg. '/people/plgtestowy' - points to your home directory
• $PLG_USER_SCRATCH_SHARED, $SCRATCH eg. '/net/scratch/people/plgtestowy/' - points to your private scratch storage - this is where data should be stored during computations

Team-wide (append team name to each variable):

• PLG_GROUPS_STORAGE, PLG_GROUPS_SHARED, eg. '/net/archive/groups' - points to where teams are expected to deposit their data (including input data, output data and data shared between team members)

Local scratch storage

For computations which require a large quantity of metadata operations (open/close/create/remove) it is possible to set up dedicated scratch storage. To do so, please supply the -C localfs parameter when submitting your job (i.e. at the resource allocation stage, before the job ID is assigned).

For example, when submitting jobs in batch mode, use the following directive: #SBATCH -C localfs

Any job submitted in this fashion will be assigned a separate scratch volume, available on the computational nodes and referenced by the $SCRATCH_LOCAL environmental variable (using the /localfs/<task_id> naming convention). $SCRATCH_LOCAL space is optimized for processing small files. Each computational node possesses a separate $SCRATCH_LOCAL volume, available only during execution of the given job.

The size of $SCRATCH_LOCAL is 768 GiB per computational node assigned to your job. Please contact the administrators if you require more scratch space.

Caution: The $SCRATCH_LOCAL volume is available only on those computational nodes which have been allocated to your job, and only while the job is executing. Furthermore, it is not shared across nodes (i.e. each node possesses a separate volume). Following completion of your job, you will lose any data saved in $SCRATCH_LOCAL - if this data is relevant, it must first be copied to your team storage ($PLG_GROUPS_STORAGE/<team_name>) or to your home directory ($HOME).

MEMFS RAM storage

In order to allocate RAM for storage of temporary files, please add the "-C memfs” parameter to your job specification. For example, use the following directive in your batch script: #SBATCH -C memfs

A storage volume will be set up for your job, referenced by the $MEMFS environmental variable. Please note that memory allocated to MEMFS storage counts towards the total memory allocated for your job and declared through "--mem" or "--mem-per-cpu".

Caution: When using MEMFS for file storage, be aware of the following limitations:

• this method is only available for single-node jobs
• the total amount of memory consumed by your job, including any MEMFS storage, must not exceed the value declared through "--mem" or "--mem-per-cpu".
• this method may only be used if the total memory requirements of your job (including MEMFS storage) do not exceed the memory available on a single node (120 GB per standard Prometheus node).
• when using MEMFS, it is recommended to request allocation of a full node for your job.

Running jobs

The SLURM scheduling system is used to manage jobs on the Prometheus cluster.

A job may be executed in batch or interactive mode.

Before running any jobs, please familiarize yourself with the grant system

Batch mode

To run a job in batch mode, use the sbatch command. Usage: sbatch script.sh.
All scheduler options should be preceded by the #SBATCH keyword (do not forget the #!).
For more information see: man sbatch and sbatch --help.

Sample script:

CAUTION! This sample script is provided as a reference and SHOULD NOT be executed as is. Before running any jobs, make sure that the scheduler options accurately reflect your requirements e.g. with regard to grant ID, partition name, execution time etc.

#!/bin/bash -l
## Job name
#SBATCH -J ADFtestjob
## Number of allocated nodes
#SBATCH -N 1
## Number of tasks per node (by default this corresponds to the number of cores allocated per node)
#SBATCH --ntasks-per-node=1
## Memory allocated per core (default is 5GB)
#SBATCH --mem-per-cpu=1GB
## Max task execution time (format is HH:MM:SS)
#SBATCH --time=01:00:00 
## Name of grant to which resource usage will be charged
#SBATCH -A <grant_id>
## Name of partition
#SBATCH -p plgrid-testing
## Name of file to which standard output will be redirected
#SBATCH --output="output.out"
## Name of file to which the standard error stream will be redirected
#SBATCH --error="error.err"

## change to sbatch working directory
cd $SLURM_SUBMIT_DIR

srun /bin/hostname
module load plgrid/apps/adf/2014.07 
adf input.adf

MPI jobs

Running MPI jobs on Prometheus is similar to running them on other SLURM clusters with integrated MPI libraries. Prometheus provides various versions of MPI libraries (mostly OpenMPI and IntelMPI) for a variety of popular compilers. When running MPI jobs, the -n and -np scheduler options should not be explicitly specified - instead, the system will select optimal values based on allocation parameters expressed in the startup script. Usage of -n and -np should be restricted to exceptional cases which call for nonstandard configurations (e.g. local threading).

CAUTION! This sample script is provided as a reference and SHOULD NOT be executed as is. Before running any jobs, make sure that the scheduler options accurately reflect your requirements e.g. with regard to grant ID, partition name, execution time etc.

#!/bin/bash -l
## Job name
#SBATCH -J MPITest
## Number of allocated nodes
#SBATCH -N 2
## Number of tasks per node (by default this corresponds to the number of cores allocated per node)
#SBATCH --ntasks-per-node=24
## Memory allocated per core (default is 5GB)
#SBATCH --mem-per-cpu=5GB
## Max task execution time (format is HH:MM:SS)
#SBATCH --time=01:00:00 
## Name of grant to which resource usage will be charged
#SBATCH -A <grant_id>
## Name of partition
#SBATCH -p plgrid-testing
## Name of file to which standard output will be redirected
#SBATCH --output="output.out"
## Name of file to which the standard error stream will be redirected
#SBATCH --error="error.err"

srun /bin/hostname

## Load the IntelMPI module
module add plgrid/tools/impi

## change to sbatch working directory
cd $SLURM_SUBMIT_DIR

mpiexec ./calcDiff 100 50

It is considered good practice to compile and execute applications in an environment comprising the same set of modules. We recommend using the mpiexec wrapper. Computational resource allocation - in this case 2 nodes, 24 cores each - results in 48 MPI tasks available for execution. It is permissible to run a series of MPI applications with a single script, but given their execution time (which is usually long) this approach is discouraged. Running multiple MPI applications in parallel within a single allocation is not allowed due to resource contention issues.

Simple parallelization of independent tasks

Parallel processing is a typical use case scenario for computing clusters. One representative example involves processing multiple images with a shared algorithm. SLURM provides a way to easily parallelize such tasks with the srun command. A sample script which runs multiple computations in parallel (depending on their specifications) is provided below.

CAUTION! This sample script is provided as a reference and SHOULD NOT be executed as is. Before running any jobs, make sure that the scheduler options accurately reflect your requirements e.g. with regard to grant ID, partition name, execution time etc.

#!/bin/bash -l
## Job name
#SBATCH -J testjob
## Number of allocated nodes
#SBATCH -N 2
## Number of tasks per node (by default this corresponds to the number of cores allocated per node)
#SBATCH --ntasks-per-node=24
## Memory allocated per core (default is 5GB)
#SBATCH --mem-per-cpu=5GB
## Max task execution time (format is HH:MM:SS)
#SBATCH --time=01:00:00 
## Name of grant to which resource usage will be charged
#SBATCH -A <grant_id>
## Name of partition
#SBATCH -p plgrid-testing
## Name of file to which standard output will be redirected
#SBATCH --output="output.out"
## Name of file to which the standard error stream will be redirected
#SBATCH --error="error.err"

module load plgrid/tools/imagemagick

## change to sbatch working directory
cd $SLURM_SUBMIT_DIR

ls *.tif | xargs -t -d "\n" -P ${SLURM_NTASKS} -n 1 srun -n 1 -N 1 --mem=5gb mogrify -format png

The above script runs the mogrify application for each *tif file in the working directory, converting the image to the png format. We begin by listing all input files, and feed the output into the xargs command, which (through the -P flag) ensures parallel execution of each subsequent srun instance. Each instance is supplied with a single input parameter, with the maximum level of parallelism defined as ${SLURM_NTASKS}. Thus, each invocation of srun runs the mogrify application on a single core, with 5GB of assigned memory.

Interactive mode

To run shell commands in interactive mode, use the following (sample) command: srun -p plgrid-testing -N 1 --ntasks-per-node=1 -n 1 -A <grant_id> --pty /bin/bash -l

This will run a job in the plgrid-testing partition, on a single node, using a single core.

The srun command is typically used to execute the specified operation on the assigned resources.
However, if resources have not been allocated in advance, srun also takes care of performing the allocation.

CAUTION: Graphic mode (using X11) is only available on the Prometheus access host.

Array jobs

Array jobs provide a way to schedule large numbers of similar coputational tasks. Each instance of an array job is assigned an index referenced by the $SLURM_ARRAY_TASK_ID environmental variable, which can be used to parameterize the given instance. Submitting array jobs is similar to submitting ordinary batch jobs. The following sample script submits an array job with the task ID range defined by the --array (or -a) scheduler option:

CAUTION! This sample script is provided as a reference and SHOULD NOT be executed as is. Before running any jobs, make sure that the scheduler options accurately reflect your requirements e.g. with regard to grant ID, partition name, execution time etc.

#!/bin/bash -l
## Job name
#SBATCH -J testjob
## Number of allocated nodes
#SBATCH -N 1
## Number of tasks per node (by default this corresponds to the number of cores allocated per node)
#SBATCH --ntasks-per-node=1
## Memory allocated per core (default is 5GB)
#SBATCH --mem-per-cpu=5GB
## Max task execution time (format is HH:MM:SS)
#SBATCH --time=01:00:00 
## Name of grant to which resource usage will be charged
#SBATCH -A <grant_id>
## Name of partition
#SBATCH -p plgrid-testing
## Name of file to which standard output will be redirected
#SBATCH --output="output.out"
## Name of file to which the standard error stream will be redirected
#SBATCH --error="error.err"
## Array index range
#SBATCH --array=0-100

## change to sbatch working directory
cd $SLURM_SUBMIT_DIR

myCalculations $SLURM_ARRAY_TASK_ID

GPGPU jobs

Accessing GPGPU nodes

A special partition named plgrid-gpu has been set up for jobs which require access to GPGPU hardware. In order to schedule such jobs, it is first necessary to obtain a grant specifically dedicated to GPGPU processing. This grant should not be used to run any other types of computational tasks.

When applying for your grant, please specify that you require access to plgrid-gpu. It is also recommended (although not strictly required) to put gpu in your grant name (e.g. gpucomputations) - this will help prevent potential mix-ups.

All applications for access to plgrid-gpu are evaluated on a case-by-case basis by the resource provider.

General information

The SLURM scheduler treats GPUs as generic resources (GRES) identified by the "gpu" codeword.
You can find out which nodes/partitions provide such resources by using the sinfo command:

sinfo -o '%P || %N || %G'

Jobs are submitted by adding --partition=plgrid-gpu --gres=gpu[:count] to your list of scheduler options. If the count argument is omitted, the system will default to allocating a single GPU adapter per processing node.

Following submission of your job, the scheduler automatically sets the $CUDA_VISIBLE_DEVICES environmental variable and permits access to allocated GPUs.

Interactive jobs

For example, to run a single-node interactive job with 2 GPUs, type:

srun -p plgrid-gpu -N 1 -n 24 -A <grant_id> --gres=gpu:2 --pty /bin/bash -l

Interactive MPI jobs

Caution: running interactive MPI jobs (e.g. for testing) is a special case. Due to the fact that SLURM treats GPUs differently than general-purpose processors, launching interactive computations calls for a nonstandard procedure.

For example, to run an interactive job on 2 nodes with 2 GPUs per node (4 GPUs total) for 1 hour, specify the following:

salloc -p plgrid-gpu -N 2 --ntasks-per-node 24 -n 48 -A <grant_id> --gres=gpu:2 --time 1:00:00

The salloc command allocates our job and returns its ID - for example, 1234.

Next, we use srun requesting 0 GPUs for the job whose ID was returned by salloc:

srun --jobid=1234 --gres=gpu:0 -O --pty /bin/bash -l

This gives us shell access on one of the computing nodes. Now, after loading the appropriate MPI module, we may run our actual application via mpirun or mpiexec. It is important to avoid exporting environmental settings - e.g., when using IntelMPI, remember to specify the -genvnone flag. The application will retain access to all GPUs allocated via salloc regardless of the gpu:0 parameter in srun.

Once the computation has concluded, the allocation should be manually cleared by typing:

scancel 1234

Batch jobs

No special procedure is required for batch jobs. For example, to request one GPU adapter per computing node, simply add the following to your script:

#SBATCH --partition=plgrid-gpu
#SBATCH --gres=gpu

Note: Further details regarding SLURM commands may be obtained by typing man sbatch

Monitoring queues, partitions, nodes, jobs and resources

• Dedicated Prometheus scripts

  • pro-jobs - lists jobs queued by or running for the current user
  • pro-jobs-history - lists completed jobs for the current user
  • pro-fs - returns disk usage statistics for the current user
  • pro-groupdir-fix - used to adjust access permissions for the team data directory
  • pro-grant-jobs - lists jobs scheduled under the user's active grants (including jobs whch belong to other users sharing these grants)

• SLURM system commands

  • squeue - list of currently queued/running jobs
  • scontrol show job [<job_ID>] - details of a given job
  • sstat -a -j <job_ID>/<job_ID.batch> - resource consumption for each step of a currently running job
  • scancel <job_ID> - cancel a queued or running job
  • sacct - resource consumption for jobs/steps which have already been completed
  • scontrol show partition [<partition_name>] - details of a given partition
  • sinfo - list of nodes
  • scontrol show node [<node_name>] - details of a given node

General rules for using the Prometheus cluster

• Any computations launched on the access host will be killed without notice. The access host should only be used to prepare input files and submitting jobs to the scheduler. Since transferring large quantities of data may cause significant load on the access host, it is recommended to allocate a computing node for this purpose (in interactive mode, e.g. srun -p plgrid -A grant_id -n 1 --pty /bin/bash -l).
• Compiling applications - this should be performed on a computing node, not on the access host (in interactive mode, e.g. srun -p plgrid -A grant_id -n 1 --pty /bin/bash -l).
• Computations which are out of scope of the submitted research agenda or which consume a large fraction of the cluster's resources will be killed without notice.
• Working datasets and intermediate results should be stored and processed in the Lustre filesystem available under $SCRATCH.
• Computations should not make use of the /tmp directory to store temporary data - instead, please use directories referenced by $SCRATCH, $SCRATCHDIR or $TMPDIR.
• The access host imposes a 6GB memory limit for all processes belonging to the current user. Exceeding this limit will cause the offending process to be automatically killed.

Contact

Assistance for PL-Grid users is available at: Helpdesk PLGrid
System administrators may be contacted by e-mail at: prometheus@cyfronet.pl, or by phone at +48 12 632 33 55 int. 603.